Arc2Face
Arc2Face is a foundation model for generating high-quality, ID-consistent human faces. It functions as an ID-conditioned system that creates realistic images of any subject using only their ArcFace embedding, typically within seconds. Built upon Stable Diffusion, it leverages extensive training on the WebFace42M dataset to achieve superior identity similarity compared to existing models. A key recent advancement is the Expression Adapter, which enables precise generation of any facial expression, including rare, asymmetric, subtle, or extreme variants, while maintaining strict identity consistency. The framework is highly extensible, supporting various input modalities through integrations like ControlNet for pose control and LCM-LoRA for accelerated inference. Developed by researchers from Imperial College London and FAU Erlangen-Nürnberg, the software is available along with its training dataset, demo spaces, and ComfyUI support. It is designed for applications requiring high-fidelity face synthesis, identi